AI control of seat belts — the failure does not show up until the day of the crash.

The problem

Four sources of failure, and every one of them lives in a different system.

A modern seat belt brings together four groups of components, and each one has its own defect pattern:

1. The webbing — woven polyester, thread marking, end stitching. A skip in the weave or a dropped stitch changes the load curve.
2. The retractor — spring, inertial sensor, pyrotechnic pretensioner. If the spring torque falls outside the window or the pretensioner carries a bad weld, nobody notices until the dynamic test (and sometimes not even there).
3. The hardware and the anchorage — stamped sheet metal, spot welding, end riveting. A weld with porosity only shows up in the destructive test, and that test is sampled.
4. The model recipe — adult webbing vs. child kit, length, color, OEM-specific connector. The SKU changeover is exactly when parts get mixed up.

The quality manager knows the dynamic bench at end of line is destructive and sampled. The static load test covers only a fraction. The classic system (bench + checklist + after-the-fact audit) works 99% of the time — and that 1% is what makes the evening news. Every mass seat belt recall started exactly like that.

How it fits the IRIS system

iLEAN does not replace the bench — it puts 100% in-line control on top of it.

The seat belt problem is not a lack of a bench: it is information split into islands (the bench result on its own PC, the stitching logged by the sewing machine, the retractor torque in another window, the model recipe in the MES). At the critical moment — the unit leaving the line — those four sources are never cross-referenced. iLEAN acts as the putty that fills those gaps without touching the bench or the sewing machine.

Edge sees the webbing, the stitching and the retractor on the line. Connect captures the bench result and the model recipe. The agent cross-references against UNECE R16 and holds. The person signs — never the other way round.

The three iLEAN pieces applied to seat belt control:

• Edge — a terminal with machine vision (CNN) over the critical stations: image of the webbing (weave defects, marking thread), of the end stitching, of the riveted hardware and of the retractor marking. It fires the actuator (light stack, diverter) in milliseconds if something does not add up. It works without a network: if the plant loses WiFi, Edge keeps reading and holding units.
• Connect — captures the load test bench result (static and dynamic) whether it comes from a modern bus, from the bench's old PC, or read by a camera off the display. It also captures the retractor spring torque from the calibration machine and the model recipe from the ERP/MES. Whatever arrives from outside (an email from the pretensioner supplier flagging a suspect batch) enters at second zero.
• Agent — cross-references the Edge image, the bench result, the retractor torque, the pyrotechnic pretensioner batch, the destination VIN and the model recipe. If something does not add up, it holds the unit and alerts the quality manager on their own channel. The person validates and signs; the line never restarts on its own.

See the full IRIS architecture →

Before and after

Sampled bench + audit vs. in-line control with iLEAN

Aspect	Today (bench + checklist + after-the-fact audit)	With iLEAN Edge + Connect + Agent
Load test coverage	Sampled by batch, destructive	Bench kept + 100% in-line vision per unit
End stitching defects	Human eye, end of shift	CNN vision over the stitching, before winding
Retractor torque	Calibrated by batch, logged locally	Captured by Connect, linked to the serial number
Multi-model SKU changeover	Paper checklist, risk at the changeover	Edge validates the fitted part vs. the model recipe
Batch-to-VIN link	Rebuilt if there is an incident, weeks	Per-unit dossier, automatic, ready for PPAP/8D
Operation without a network	n/a	Edge keeps running on the panel's own light

Impact estimate

Impact estimate for your plant — to be validated with your numbers.

The block below is an estimate to be validated with the specific data of your plant. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Multi-model seat belt assembly line (adult, child kit) with an end-stitching sewing machine, retractor calibration and a dynamic load bench at the end.
• Edge pilot on one line (camera over the stitching + camera over the retractor + integration with the bench). First value expected within a few weeks.
• Indicative payback between 4 and 9 months, depending on how often the OEM has raised PRR/8D in recent years.
• The hard lever is a single recall avoided: the cost of one passive safety workshop callback pays for the pilot and the rollout to the other lines.

And the quality manager's reasonable doubt

“What if the AI gets it wrong and lets a belt with defective stitching through?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely compares an image of the stitching against a learned pattern, the best models brought error below 1.5% ^[1]. And even so, what is critical is never decided alone: iLEAN holds the unit and the person signs. The automotive quality standard of the order of 25 PPM ^[2] is not sustained with sampling: it is sustained with vision on every part.

^[1] OpenAI paper “Why Language Models Hallucinate”, 2025 — on the reliability of AI in anchored tasks.

^[2] Symestic — automotive quality standard of the order of 25 PPM.